RAG Interview Questions, issue 25, Jul 29, 2026
The "Lost in the Middle" Trap
Senior ML Engineer interview at Anthropic, and the interviewer asks:
“Your retriever pulls top-10 chunks, reranks them, and you feed all 10 to the LLM to be safe. Walk me through exactly what breaks, and when you’d retrieve fewer documents on purpose.”
Don’t say: “More context gives the model more to work with, so top-10 is safer than top-3.”
Why a massive 128K context window won't save your RAG pipeline from U-shaped attention decay, and how head-tail reordering rescues your most relevant data from the dead zone.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.